OpenCV multiple template matching - Improving matching-CodePudding

I try to recognize 4 the same fiducial marks on a map. With help of the internet, I have created something, but I am looking for ways to improve the search, since the result is far from perfect.

What I tried so far:

Changing the threshold
Trying different cv2 methods
Making the image and the template smaller

This is my code:

import cv2
import numpy as np
from imutils.object_detection import non_max_suppression
  
# Reading and resizing the image

big_image = cv2.imread('20221028_093830.jpg')
 
scale_percent = 10 # percent of original size
width = int(big_image.shape[1] * scale_percent / 100)
height = int(big_image.shape[0] * scale_percent / 100)
dim = (width, height)

img = cv2.resize(big_image, dim, interpolation = cv2.INTER_AREA)


temp = cv2.imread('try_fiduc.png')
  
# save the image dimensions
W, H = temp.shape[:2]
  
# Converting them to grayscale
img_gray = cv2.cvtColor(img, 
                        cv2.COLOR_BGR2GRAY)
temp_gray = cv2.cvtColor(temp,
                         cv2.COLOR_BGR2GRAY)

# Blur the image
img_blurred = cv2.GaussianBlur(img_gray, (7, 7), 0)

# Increasing contrast
img_contrast = img_blurred*3

# Passing the image to matchTemplate method
match = cv2.matchTemplate(
    image=img_contrast, templ=temp_gray, 
  method=cv2.TM_CCOEFF)\

# Define a minimum threshold
thresh = 6000000

# Select rectangles with confidence greater than threshold
(y_points, x_points) = np.where(match >= thresh)
  
# initialize our list of rectangles
boxes = list()
  
# loop over the starting (x, y)-coordinates again
for (x, y) in zip(x_points, y_points):
    
    # update our list of rectangles
    boxes.append((x, y, x   W, y   H))
  
# apply non-maxima suppression to the rectangles
# this will create a single bounding box
boxes = non_max_suppression(np.array(boxes))
  
# loop over the final bounding boxes
for (x1, y1, x2, y2) in boxes:
    
    # draw the bounding box on the image
    cv2.rectangle(img, (x1, y1), (x2, y2),
                  (255, 0, 0), 3)
  

# Show the template and the final output
cv2.imshow("Template", temp_gray)
cv2.imshow("Image", img_contrast)
cv2.imshow("After NMS", img)
cv2.waitKey(0)
  
# destroy all the windows manually to be on the safe side
cv2.destroyAllWindows()

This is my template:

This is my image:

What are more ways to improve the template matching? In the end I want to be able to recognize them from further distance, and not have the false match. Any help would be appreciated.

CodePudding user response：

Here is how I would do that in Python/OpenCV. Mostly the same as yours with several changes.

First, I would not bother computing the dim for resize. I would just use your scale_percent/100 so a fraction. Resize permits that in place of the size.

Second, I would threshold your images and invert the template so that you are matching black rings in both the image and template.

Third, I would use TM_SQDIFF and find values below a threshold.

import cv2
import numpy as np
from imutils.object_detection import non_max_suppression
  
# Reading and resizing the image

big_image = cv2.imread('diagram.jpg')
 
scale_percent = 10 # percent of original size
scale = scale_percent/100

img = cv2.resize(big_image, (0,0), fx=scale, fy=scale, interpolation = cv2.INTER_AREA)

temp = cv2.imread('ring.png')

# save the image dimensions
W, H = temp.shape[:2]
  
# Converting them to grayscale
img_gray = cv2.cvtColor(img, 
                        cv2.COLOR_BGR2GRAY)
temp_gray = cv2.cvtColor(temp,
                         cv2.COLOR_BGR2GRAY)

# threshold (and invert template)
img_thresh = cv2.threshold(img_gray, 0, 255, cv2.THRESH_BINARY cv2.THRESH_OTSU)[1]
temp_thresh = cv2.threshold(temp_gray, 0, 255, cv2.THRESH_BINARY_INV cv2.THRESH_OTSU)[1]

# Passing the image to matchTemplate method
match = cv2.matchTemplate(
    image=img_thresh, templ=temp_thresh, 
  method=cv2.TM_SQDIFF)\

min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(match)
print(min_val, max_val)

# Define a threshold
# thresh between 40000000 and 60000000 works
thresh = 50000000

# Select rectangles with confidence less than threshold for TM_SQDIFF
(y_points, x_points) = np.where(match <= thresh)
  
# initialize our list of rectangles
boxes = list()
  
# loop over the starting (x, y)-coordinates again
for (x, y) in zip(x_points, y_points):    
    # update our list of rectangles
    boxes.append((x, y, x   W, y   H))
  
# apply non-maxima suppression to the rectangles
# this will create a single bounding box
boxes = non_max_suppression(np.array(boxes))
  
# loop over the final bounding boxes
result = img.copy()
for (x1, y1, x2, y2) in boxes:    
    # draw the bounding box on the image
    cv2.rectangle(result, (x1, y1), (x2, y2),
                  (255, 0, 0), 3)

# save result
cv2.imwrite('diagram_match_locations.jpg', result) 

# Show the template and the final output
cv2.imshow("Template_thresh", temp_thresh)
cv2.imshow("Image_thresh", img_thresh)
cv2.imshow("After NMS", result)
cv2.waitKey(0)
  
# destroy all the windows manually to be on the safe side
cv2.destroyAllWindows()

Result: